v3.11.5: token-aware compaction, vision filter, universal adaptive compaction, smart-continue text detection

CHANGELOG.md

@@ -1,5 +1,29 @@
 # Changelog
 
+## v3.11.5 (2026-05-26)
+
+**Vision Filter, Token-Aware Compaction, Universal Adaptive Compaction, Smart-Continue Text Detection**
+
+### Critical Fixes
+
+- **Token-aware compaction for small-context models (FIX)**: `_crof_compact_for_retry()` had an early return at `len(input_data) <= limit` (item count) — if you had 25 items × 1600 tokens = 40K tokens, it skipped compaction entirely because 25 < 30 (the default item limit). Now also checks estimated token count vs learned model max, and compacts when either item count OR token count exceeds limits. Fixes repeated `context_length_exceeded` errors on models like 0G-GLM-5.1 (~35K token context).
+- **Proactive compaction now token-aware**: Previously only triggered when item count > 30. Now also triggers when estimated tokens exceed 80% of the model's learned token limit, even if item count is below the threshold. Prevents the first-request failure pattern on small-context models.
+- **Compaction aggression threshold**: Changed `est > max_tok` to `est >= max_tok * 0.9` to avoid edge case where estimated tokens exactly equal the limit and compaction is skipped.
+- **Removed all `crof.ai` gates from adaptive compaction**: Proactive compaction, `finish_reason=length` retry, `_crof_record`, and compaction logging were gated behind `"crof.ai" in TARGET_URL`. These gates prevented OpenAdapter and other providers from getting proactive/retry compaction, causing repeated `context_length_exceeded` failures. Now applies universally to ALL providers.
+
+### New Features
+
+- **Vision model detection + image stripping**: `_strip_images_from_input()` and `_model_supports_vision()` detect vision capability by model name pattern. Non-vision models (deepseek, glm, mixtral, llama, command, dbrx, qwen, phi-3) have `input_image`/`image_url` parts stripped and replaced with `[User attached image: filename — this model does not support vision]` text notice. Vision models (gpt-4o, gemini, claude, qwen-vl, glm-5v) keep images intact. Applied in 3 paths: main request, context_length_exceeded retry, smart-continue nudge.
+- **Token estimation and per-model limit learning**: `_estimate_tokens()`, `_estimate_input_tokens()`, `_get_model_max_tokens()`, `_set_model_max_tokens()`. Extracts `~N tokens` from `context_length_exceeded` error messages and stores per-model token limits. Used by proactive compaction and retry compaction to adjust `keep` count dynamically.
+- **Compaction aggression levels**: `_crof_compact_for_retry()` accepts `aggression` parameter (0=normal, 1=extreme). Extreme mode kicks in when estimated tokens > 1.5× the learned limit or on 2nd+ retry attempt. Reduces `keep` count to minimum, ensuring the compacted request fits within model limits.
+- **Smart-continue text-tool detection**: Removed hard requirement for `has_function_call_output(input_data)`. Added `_TOOL_CALL_TEXT_PATTERNS` and `_text_looks_like_tool_calls()` to trigger nudging when model outputs text matching tool-call patterns (e.g., `• (exec_command cmd ...)`, `write_to_file`, `exec_command`) even without prior `function_call_output` in context. Essential for models like 0G-GLM-5.1 that never emit real `function_call_output` items.
+- **Parenthesized tool call regex**: `_PAREN_TC_RE` pattern to match `• (name args...)` format from non-vision models that output tool calls as parenthesized text.
+
+### GUI Fixes
+
+- **Active endpoint sync**: Added `set_active_endpoint()` and `validate_active_endpoint()` to Linux GTK GUI. Syncs `.active-endpoint.json` with `config.toml` on every launch; auto-removes stale references to deleted providers. Fixed `"Error loading configuration: No such file or directory (os error 2)"` crash when active endpoint referenced a deleted provider.
+- **Config state**: `~/.codex/.active-endpoint.json` and `config.toml` model catalog path validated and auto-corrected on GUI startup.
+
 ## v3.11.0 (2026-05-26)
 
 **Cobra PR Merge + Smart Continuation + API Key Hot-Reload**

README.md

@@ -130,6 +130,10 @@ A three-component system:
 - **Response store TTL** — evicts stored responses older than 10 minutes, prevents memory leaks
 - **Bounded stream buffers** — 8MB cap prevents OOM on pathological responses
 - **Dual logging** — all proxy messages written to both stderr and `~/.cache/codex-proxy/proxy.log`
+- **Vision model detection** (v3.11.5) — automatically strips images for non-vision models (DeepSeek, GLM, Qwen, etc.) and replaces with text notice; vision-capable models (GPT-4o, Gemini, Claude, Qwen-VL) keep images intact
+- **Token-aware compaction** (v3.11.5) — learns per-model token limits from `context_length_exceeded` errors; proactively compacts when estimated tokens exceed 80% of limit; prevents repeated context overflow on small-context models (~35K tokens)
+- **Universal adaptive compaction** (v3.11.5) — compaction now works for ALL providers (was Crof.ai-only); proactive + retry compaction with aggression levels (normal/extreme)
+- **Smart-continue text detection** (v3.11.5) — triggers continuation nudging when model outputs text matching tool-call patterns, essential for text-only models that never emit real `function_call_output` items
 - Zero dependencies — pure Python stdlib
 
 ### Command Code Adapter

install.sh

@@ -3,13 +3,13 @@ set -e
 
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 
-if [ -f "$SCRIPT_DIR/codex-launcher_3.11.0_all.deb" ]; then
+if [ -f "$SCRIPT_DIR/codex-launcher_3.11.5_all.deb" ]; then
-    echo "Installing codex-launcher_3.11.0_all.deb ..."
+    echo "Installing codex-launcher_3.11.5_all.deb ..."
-    sudo dpkg -i "$SCRIPT_DIR/codex-launcher_3.11.0_all.deb"
+    sudo dpkg -i "$SCRIPT_DIR/codex-launcher_3.11.5_all.deb"
 else
-    echo "WARNING: codex-launcher_3.11.0_all.deb not found; copying files manually."
+    echo "WARNING: codex-launcher_3.11.5_all.deb not found; copying files manually."
 fi
-echo "Installed v3.11.0 via .deb package."
+echo "Installed v3.11.5 via .deb package."
     echo "  translate-proxy.py   -> /usr/bin/translate-proxy.py"
     echo "  codex-launcher-gui   -> /usr/bin/codex-launcher-gui"
     echo "  cleanup-codex-stale  -> /usr/bin/cleanup-codex-stale.sh"

src/codex-launcher-gui

@@ -20,12 +20,22 @@ BGP_POOLS_FILE = HOME / ".codex/bgp-pools.json"
 LOG_DIR = HOME / ".cache/codex-desktop"
 LAUNCH_LOG = LOG_DIR / "launcher.log"
 PROXY_CONFIG_DIR = HOME / ".cache/codex-proxy"
+ACTIVE_ENDPOINT_FILE = HOME / ".codex/.active-endpoint.json"
 DEFAULT_CONFIG = """model = ""
 model_provider = ""
 model_catalog_json = ""
 """
 
 CHANGELOG = [
+    ("3.11.5", "2026-05-26", [
+        "Token-aware compaction: fixes context_length_exceeded on small-context models",
+        "Proactive compaction triggers on token count, not just item count",
+        "Universal adaptive compaction for all providers (removed crof.ai gates)",
+        "Vision model detection + image stripping for non-vision models",
+        "Per-model token limit learning from error messages",
+        "Smart-continue text-tool detection for text-only models",
+        "Active endpoint sync: auto-removes stale references on startup",
+    ]),
     ("3.11.0", "2026-05-26", [
         "Merge cobra PR: concurrency semaphore (max 3), auto-continue for truncated text",
         "SO_REUSEADDR on sticky port, proxy-stderr.log, stream diagnostics logging",
@@ -33,7 +43,7 @@ CHANGELOG = [
         "Restart Proxy button: only restarts proxy without killing Codex Desktop",
         "Tool call argument normalizer: fixes Arguments→arguments, strips markdown wrapping",
         "Smart-continue loop (2× retries): escalating nudges when model stops text-only mid-task",
-        "XML tool call extraction: parses <tool_call> patterns from text, injects as real calls",
+        "XML tool call extraction: parses <name> patterns from text, injects as real calls",
         "Auto-continue + smart-continue ordered with skip guard to avoid double-firing",
         "API key hot-reload with mtime tracking + /admin/reload + /admin/verify-key endpoints",
         "GUI hot-reload: auto-refreshes proxy key on endpoint edit, verifies with upstream",
@@ -923,6 +933,27 @@ def restore_config():
         shutil.copy2(str(CONFIG_BAK), str(tmp))
         os.replace(str(tmp), str(CONFIG))
 
+def set_active_endpoint(name):
+    ACTIVE_ENDPOINT_FILE.parent.mkdir(parents=True, exist_ok=True)
+    write_secure_text(ACTIVE_ENDPOINT_FILE, json.dumps({"active": name}, indent=2))
+
+def validate_active_endpoint(logfn=None):
+    if not ACTIVE_ENDPOINT_FILE.exists():
+        return
+    try:
+        d = json.loads(ACTIVE_ENDPOINT_FILE.read_text())
+        active = d.get("active", "")
+        if not active:
+            return
+        eps = load_endpoints()
+        names = {ep.get("name", "") for ep in eps}
+        if active not in names:
+            ACTIVE_ENDPOINT_FILE.unlink()
+            if logfn:
+                logfn(f"Removed stale active-endpoint '{active}' (provider no longer exists)")
+    except Exception:
+        pass
+
 def write_secure_text(path, text):
     path.parent.mkdir(parents=True, exist_ok=True)
     tmp = path.with_suffix(path.suffix + ".tmp")
@@ -1862,6 +1893,7 @@ class LauncherWin(Gtk.Window):
         self._proc = None
         self._endpoints_data = load_endpoints()
         recover_config_if_needed()
+        validate_active_endpoint()
 
         vbox = Gtk.Box(orientation=Gtk.Orientation.VERTICAL, spacing=8)
         self.add(vbox)
@@ -2607,6 +2639,8 @@ class LauncherWin(Gtk.Window):
                 begin_config_transaction(f"launch:{ep['name']}")
                 write_config_for_native(ep, model)
 
+            set_active_endpoint(ep["name"])
+
             if target == "desktop":
                 if needs_proxy:
                     _kill_existing_desktop(self.log)
@@ -2664,6 +2698,7 @@ class LauncherWin(Gtk.Window):
 
             begin_config_transaction(f"launch:bgp:{pool['name']}")
             write_config_for_translated(bgp_ep, model, port)
+            set_active_endpoint(pool["name"])
 
             if target == "desktop":
                 _kill_existing_desktop(self.log)

src/codex_launcher_lib.py

@@ -83,14 +83,24 @@ model_catalog_json = ""
 """
 
 CHANGELOG = [
+    ("3.11.5", "2026-05-26", [
+        "Token-aware compaction: fixes context_length_exceeded on small-context models (25 items × 1600 tokens)",
+        "Proactive compaction triggers on token count (>80% model limit), not just item count",
+        "Universal adaptive compaction: removed crof.ai-only gates, all providers get compaction",
+        "Vision model detection: strips images for non-vision models, keeps for vision-capable ones",
+        "Per-model token limit learning from context_length_exceeded error messages",
+        "Compaction aggression levels: normal vs extreme when tokens > 1.5× model limit",
+        "Smart-continue text-tool detection: triggers on tool-call text patterns, not just function_call_output",
+        "Active endpoint sync: GUI auto-removes stale endpoint references on startup",
+    ]),
     ("3.11.0", "2026-05-26", [
         "Merge cobra PR: concurrency semaphore (max 3), auto-continue for truncated text",
         "SO_REUSEADDR on sticky port, proxy-stderr.log, stream diagnostics logging",
         "Timeout/OSError handler sends response.failed SSE instead of silent drop",
         "Restart Proxy button: only restarts proxy without killing Codex Desktop",
-        "Tool call argument normalizer: fixes Arguments→arguments, strips markdown wrapping",
+        "Tool call argument normalizer: fixes Arguments->arguments, strips markdown wrapping",
-        "Smart-continue loop (2× retries): escalating nudges when model stops text-only mid-task",
+        "Smart-continue loop (2x retries): escalating nudges when model stops text-only mid-task",
-        "XML tool call extraction: parses <tool_call> patterns from text, injects as real calls",
+        "XML tool call extraction: parses patterns from text, injects as real calls",
         "Auto-continue + smart-continue ordered with skip guard to avoid double-firing",
         "API key hot-reload with mtime tracking + /admin/reload + /admin/verify-key endpoints",
         "GUI hot-reload: auto-refreshes proxy key on endpoint edit, verifies with upstream",

src/translate-proxy.py

@@ -1469,6 +1469,53 @@ _CROF_ADAPTIVE = {
     "min_keep_recent": 6,
 }
 
+_model_max_tokens = {}
+_model_max_tokens_lock = threading.Lock()
+
+def _estimate_tokens(item):
+    if not isinstance(item, dict):
+        return 4
+    t = item.get("type", "")
+    if t == "message":
+        content = item.get("content", "")
+        if isinstance(content, str):
+            return max(4, len(content) // 4)
+        elif isinstance(content, list):
+            total = 4
+            for part in content:
+                pt = part.get("type", "")
+                if pt in ("input_text", "output_text"):
+                    total += max(4, len(part.get("text", "")) // 4)
+                elif pt == "input_image":
+                    total += 800
+                elif pt in ("function_call",):
+                    total += max(20, len(part.get("arguments", "{}")) // 2)
+                elif pt == "function_call_output":
+                    total += max(8, len(part.get("output", "")) // 4)
+            return total
+    elif t in ("function_call_output",):
+        return max(8, len(item.get("output", "")) // 4)
+    elif t == "function_call":
+        return max(20, len(item.get("arguments", "{}")) // 2)
+    return 4
+
+def _estimate_input_tokens(input_data):
+    if not isinstance(input_data, list):
+        return 0
+    return sum(_estimate_tokens(i) for i in input_data)
+
+def _get_model_max_tokens(model):
+    with _model_max_tokens_lock:
+        return _model_max_tokens.get(model)
+
+def _set_model_max_tokens(model, tokens):
+    if model and tokens:
+        with _model_max_tokens_lock:
+            existing = _model_max_tokens.get(model)
+            if existing is None or tokens < existing:
+                _model_max_tokens[model] = tokens
+                print(f"[ctx-limit] learned {model} max ~{tokens} tokens", file=sys.stderr)
+
 _BGP_STATS_PATH = os.path.join(_LOG_DIR, "bgp-route-stats.json")
 _bgp_stats_lock = threading.Lock()
 
@@ -1534,8 +1581,6 @@ def _sorted_bgp_routes():
     return sorted(BGP_ROUTES, key=lambda r: _score_route(r, stats))
 
 def _crof_record(model, n_items, success):
-    if TARGET_URL and "crof.ai" not in TARGET_URL:
-        return
     if not isinstance(n_items, int) or n_items < 1:
         return
     entry = {"model": model, "items": n_items, "ok": success}
@@ -1561,20 +1606,36 @@ def _crof_record(model, n_items, success):
             global_limit = v["limit"]
     _CROF_ADAPTIVE["global_item_limit"] = global_limit
 
-    if TARGET_URL and "crof.ai" in TARGET_URL:
+    print(f"[crof-adaptive] model={model} items={n_items} {'OK' if success else 'FAIL'} -> limit={ml.get('limit',30)} global={global_limit}", file=sys.stderr)
-        print(f"[crof-adaptive] model={model} items={n_items} {'OK' if success else 'FAIL'} -> limit={ml.get('limit',30)} global={global_limit}", file=sys.stderr)
 
 def _crof_item_limit(model):
     ml = _CROF_ADAPTIVE["model_limits"].get(model, {})
     per_model = ml.get("limit", 30)
     return min(per_model, _CROF_ADAPTIVE["global_item_limit"])
 
-def _crof_compact_for_retry(input_data, model):
+def _crof_compact_for_retry(input_data, model, aggression=0):
     limit = _crof_item_limit(model)
-    if not isinstance(input_data, list) or len(input_data) <= limit:
+    if not isinstance(input_data, list) or len(input_data) < 2:
+        return input_data
+
+    max_tok = _get_model_max_tokens(model)
+    est = _estimate_input_tokens(input_data)
+    over_item_limit = len(input_data) > limit
+    over_token_limit = max_tok and est >= max_tok * 0.9
+
+    if not over_item_limit and not over_token_limit:
         return input_data
 
     keep = max(_CROF_ADAPTIVE["min_keep_recent"], limit // 3)
+    if over_token_limit:
+        ratio = est / max_tok
+        if aggression >= 1 or ratio > 1.5:
+            keep = max(2, _CROF_ADAPTIVE["min_keep_recent"] // 2)
+        elif ratio > 1.2:
+            keep = max(3, keep // 2)
+        print(f"[ctx-limit] model={model} est={est}tok max={max_tok}tok ratio={ratio:.2f} -> keep={keep}", file=sys.stderr)
+    elif over_item_limit:
+        keep = max(keep, 6)
     head_end = 0
     for i, item in enumerate(input_data):
         t = item.get("type")
@@ -1607,8 +1668,7 @@ def _crof_compact_for_retry(input_data, model):
         summary_lines.append(_item_summary(item, max_len=120))
 
     summary_msg = {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "\n".join(summary_lines)}]}
-    if TARGET_URL and "crof.ai" in TARGET_URL:
+    print(f"[crof-adaptive] RETRY compact: {len(input_data)} -> {len(head)+1+len(tail)} (limit={limit}, keep={len(tail)}, agg={aggression})", file=sys.stderr)
-        print(f"[crof-adaptive] RETRY compact: {len(input_data)} -> {len(head)+1+len(tail)} (limit={limit}, keep={len(tail)})", file=sys.stderr)
     return head + [summary_msg] + tail
 
 def _item_summary(item, max_len=200):
@@ -2051,6 +2111,18 @@ def synthesize_tool_results_for_chat(input_items):
 def has_function_call_output(input_items):
     return isinstance(input_items, list) and any(i.get("type") == "function_call_output" for i in input_items)
 
+_TOOL_CALL_TEXT_PATTERNS = re.compile(
+    r'(?:^|\n)[\s•\-\*]*\(?'
+    r'(?:exec_command|write_to_file|exec_bash|bash|run_command|shell|edit_file|read_file|search_files|list_files)'
+    r'[\s:]',
+    re.I | re.MULTILINE
+)
+
+def _text_looks_like_tool_calls(text):
+    if not text or len(text) < 6:
+        return False
+    return bool(_TOOL_CALL_TEXT_PATTERNS.search(text))
+
 # ═══════════════════════════════════════════════════════════════════
 # Log redaction
 # ═══════════════════════════════════════════════════════════════════
@@ -2233,9 +2305,14 @@ def _normalize_tool_args(raw_args):
     except json.JSONDecodeError:
         return raw_args
 
-_XML_TC_RE = re.compile(r'<tool_call>(\w+)(.*?)</tool_call>', re.DOTALL)
+_XML_TC_RE = re.compile(r'exec_command(.*?)</invoke>', re.DOTALL)
 _XML_ARG_VALUE_RE = re.compile(r'</?arg_value>\s*')
 
+_PAREN_TC_RE = re.compile(
+    r'(?:^|[\n•\-\*]\s*)\(\s*(exec_command|write_to_file|exec_bash|bash|run_command|shell|edit_file|read_file|search_files|list_files)\b\s*(.*?)\)',
+    re.DOTALL | re.I
+)
+
 def _extract_xml_tool_calls(text):
     if not text:
         return []
@@ -2262,6 +2339,68 @@ def _extract_xml_tool_calls(text):
         results.append({"name": name, "args": args_str, "call_id": f"xml_{len(results)}"})
     return results
 
+_NON_VISION_MODEL_PATTERNS = re.compile(
+    r'\b(deepseek|glm|mixtral|llama\b(?!.*vision)|command|dbrx|qwen\b(?!.*vl)|phi-?3(?!.*vision))',
+    re.I
+)
+
+_vision_fail_cache = set()
+_vision_fail_lock = threading.Lock()
+
+def _model_supports_vision(model):
+    if not model:
+        return True
+    with _vision_fail_lock:
+        if model in _vision_fail_cache:
+            return False
+    if _NON_VISION_MODEL_PATTERNS.search(model):
+        return False
+    return True
+
+def _mark_vision_fail(model):
+    if model:
+        with _vision_fail_lock:
+            _vision_fail_cache.add(model)
+
+def _strip_images_from_input(input_data, model):
+    if not isinstance(input_data, list) or _model_supports_vision(model):
+        return input_data
+    modified = False
+    result = []
+    for item in input_data:
+        if item.get("type") != "message":
+            result.append(item)
+            continue
+        content = item.get("content", [])
+        if isinstance(content, str):
+            result.append(item)
+            continue
+        new_content = []
+        has_img = False
+        for part in content:
+            if isinstance(part, str):
+                new_content.append(part)
+                continue
+            pt = part.get("type", "")
+            if pt in ("input_image", "image_url"):
+                if not has_img:
+                    fname = part.get("image_url", {}).get("url", part.get("url", "image.png"))
+                    if fname.startswith("data:"):
+                        fname = "screenshot.png"
+                    new_content.append({"type": "output_text", "text": f"[User attached image: {fname} — this model does not support vision]"})
+                    has_img = True
+                    modified = True
+            else:
+                new_content.append(part)
+        if modified:
+            result.append({**item, "content": new_content})
+        else:
+            result.append(item)
+    if modified:
+        print(f"[vision-filter] stripped {sum(1 for i in input_data if i.get('type')=='message' and any(c.get('type') in ('input_image','image_url') for c in (i.get('content') or []) if isinstance(c,dict)))} images for model={model}", file=sys.stderr)
+        return result
+    return input_data
+
 def oa_input_to_messages(input_data):
     msgs = []
     tool_name_by_id = {}
@@ -4889,12 +5028,25 @@ class Handler(http.server.BaseHTTPRequestHandler):
             body["input"] = input_data
 
         crof_limit = _crof_item_limit(model)
-        _crof_eligible = TARGET_URL and "crof.ai" in TARGET_URL
+        _crof_eligible = True
-        if _crof_eligible and not compacted and isinstance(input_data, list) and len(input_data) > crof_limit:
+        if _crof_eligible and not compacted and isinstance(input_data, list):
-            print(f"[crof-adaptive] proactive compact: {len(input_data)} items > limit {crof_limit}", file=sys.stderr)
+            _needs_compact = len(input_data) > crof_limit
-            input_data = _crof_compact_for_retry(input_data, model)
+            max_tok = _get_model_max_tokens(model)
-            body = dict(body)
+            est_tok = _estimate_input_tokens(input_data) if max_tok else 0
-            body["input"] = input_data
+            if not _needs_compact and max_tok and est_tok > max_tok * 0.8:
+                _needs_compact = True
+            if _needs_compact:
+                _agg = 0
+                if max_tok and est_tok > max_tok:
+                    _agg = 1
+                print(f"[crof-adaptive] proactive compact: {len(input_data)} items, est={est_tok}tok max={max_tok}tok agg={_agg}", file=sys.stderr)
+                input_data = _crof_compact_for_retry(input_data, model, aggression=_agg)
+                body = dict(body)
+                body["input"] = input_data
+
+        # Strip images for non-vision models
+        input_data = _strip_images_from_input(input_data, model)
+        body["input"] = input_data
 
         messages = oa_input_to_messages(input_data)
         messages = _inject_stored_reasoning(messages)
@@ -4927,14 +5079,19 @@ class Handler(http.server.BaseHTTPRequestHandler):
                 except urllib.error.HTTPError as e:
                     err_body = e.read().decode()
                     if "context_length_exceeded" in err_body and attempt < max_retries:
-                        print(f"[{self._session_id}] context_length_exceeded (attempt {attempt+1}/{max_retries}), retrying with extreme compaction!", file=sys.stderr)
+                        import re as _re
+                        _tok_m = _re.search(r'~?(\d+)\s*tokens', err_body)
+                        if _tok_m:
+                            _set_model_max_tokens(model, int(_tok_m.group(1)))
+                        print(f"[{self._session_id}] context_length_exceeded (attempt {attempt+1}/{max_retries}), retrying with compaction (agg={attempt})!", file=sys.stderr)
                         policy = provider_policy()
                         if isinstance(input_data, list):
-                            print(f"[{self._session_id}] applying extreme compaction to {len(input_data)} items", file=sys.stderr)
+                            est = _estimate_input_tokens(input_data)
-                            input_data = _crof_compact_for_retry(input_data, model)
+                            print(f"[{self._session_id}] applying compaction to {len(input_data)} items ~{est}tok", file=sys.stderr)
+                            input_data = _crof_compact_for_retry(input_data, model, aggression=attempt)
                             body = dict(body)
                             body["input"] = input_data
-                            messages = oa_input_to_messages(input_data)
+                            messages = oa_input_to_messages(_strip_images_from_input(input_data, model))
                             messages = _inject_stored_reasoning(messages)
                             instructions = body.get("instructions", "").strip()
                             if instructions:
@@ -5725,9 +5882,11 @@ class Handler(http.server.BaseHTTPRequestHandler):
             last_status = None
             finish_reason = None
             has_content = False
+            has_message = False
+            has_tool_call = False
 
             def _observe_event(event):
-                nonlocal last_resp_id, last_output, last_status, finish_reason, has_content
+                nonlocal last_resp_id, last_output, last_status, finish_reason, has_content, has_message, has_tool_call
                 for line in event.strip().split("\n"):
                     if line.startswith("data: "):
                         try:
@@ -5737,7 +5896,9 @@ class Handler(http.server.BaseHTTPRequestHandler):
                                 last_output = d.get("response", {}).get("output", [])
                                 last_status = d.get("response", {}).get("status")
                                 finish_reason = "length" if last_status == "incomplete" else "stop"
-                                has_content = any(o.get("type") == "message" for o in (last_output or []))
+                                has_tool_call = any(o.get("type") == "function_call" for o in (last_output or []))
+                                has_message = any(o.get("type") == "message" for o in (last_output or []))
+                                has_content = has_message or has_tool_call
                         except Exception:
                             pass
 
@@ -5749,7 +5910,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         break
                     collected_events.append(event)
                     _observe_event(event)
-                print(f"[{self._session_id}] stream ended: events={len(collected_events)} finish={finish_reason} has_content={has_content} elapsed={time.time()-t0:.1f}s", file=sys.stderr)
+                print(f"[{self._session_id}] stream ended: events={len(collected_events)} finish={finish_reason} has_content={has_content} has_message={has_message} has_tool_call={has_tool_call} elapsed={time.time()-t0:.1f}s", file=sys.stderr)
             except (ConnectionResetError, BrokenPipeError, ConnectionAbortedError):
                 print("[translate-proxy] client disconnected during stream", file=sys.stderr)
                 _crof_record(model, n_items, False)
@@ -5805,6 +5966,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)
@@ -5813,7 +5976,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         print(f"[provider-sensor] synthetic retry failed: {e}", file=sys.stderr)
 
             # Auto-retry on finish_reason=length with no content due to too much context.
-            if finish_reason == "length" and not has_content and isinstance(input_data, list) and len(input_data) > 5 and TARGET_URL and "crof.ai" in TARGET_URL:
+            if finish_reason == "length" and not has_content and isinstance(input_data, list) and len(input_data) > 5:
                 print(f"[crof-adaptive] RETRY: finish_reason=length with no content, compacting {n_items} items", file=sys.stderr)
                 new_input = _crof_compact_for_retry(input_data, model)
                 if len(new_input) < len(input_data):
@@ -5836,6 +5999,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)
@@ -5943,9 +6108,17 @@ class Handler(http.server.BaseHTTPRequestHandler):
                 _smart_attempt = 0
                 while _smart_attempt < _smart_max:
                     _has_tool_calls_in_output = any(o.get("type") == "function_call" for o in (last_output or []))
+                    last_text = ""
+                    for o in (last_output or []):
+                        if o.get("type") == "message":
+                            for c in (o.get("content") or []):
+                                if isinstance(c, dict) and c.get("type") == "output_text":
+                                    last_text += c.get("text", "")
+                    _looks_like_tools = _text_looks_like_tool_calls(last_text)
+                    _has_prior_tool_ctx = has_function_call_output(input_data)
                     if not (finish_reason == "stop" and has_content and not _has_tool_calls_in_output
                             and isinstance(input_data, list) and len(input_data) >= 3
-                            and has_function_call_output(input_data)):
+                            and (_has_prior_tool_ctx or _looks_like_tools)):
                         break
                     _smart_attempt += 1
                     _nudges = [
@@ -5954,12 +6127,6 @@ class Handler(http.server.BaseHTTPRequestHandler):
                     ]
                     nudge_text = _nudges[min(_smart_attempt - 1, len(_nudges) - 1)]
                     # Try extracting XML tool calls from text as fallback before nudging
-                    last_text = ""
-                    for o in (last_output or []):
-                        if o.get("type") == "message":
-                            for c in (o.get("content") or []):
-                                if isinstance(c, dict) and c.get("type") == "output_text":
-                                    last_text += c.get("text", "")
                     xml_fc = _extract_xml_tool_calls(last_text)
                     if xml_fc:
                         print(f"[{self._session_id}] [smart-continue] extracted {len(xml_fc)} XML tool calls from text, injecting and retrying", file=sys.stderr)
@@ -5979,6 +6146,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                             last_resp_id = last_output = last_status = None
                             finish_reason = None
                             has_content = False
+                            has_message = False
+                            has_tool_call = False
                             for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                                 collected_events.append(event)
                                 _observe_event(event)
@@ -5988,19 +6157,21 @@ class Handler(http.server.BaseHTTPRequestHandler):
                             print(f"[{self._session_id}] [smart-continue] XML injection retry failed: {e}", file=sys.stderr)
                             break
                     _nudge_msg = {"role": "user", "content": nudge_text}
-                    nudge_messages = oa_input_to_messages(input_data) + [_nudge_msg]
+                    nudge_messages = oa_input_to_messages(_strip_images_from_input(input_data, model)) + [_nudge_msg]
                     instructions = body.get("instructions", "").strip()
                     if instructions:
                         nudge_messages.insert(0, {"role": "system", "content": instructions})
                     nudge_chat_body = self._build_chat_body(model, nudge_messages, body, stream)
                     nudge_req = urllib.request.Request(target, data=json.dumps(nudge_chat_body).encode(), headers=fwd)
-                    print(f"[{self._session_id}] [smart-continue] attempt {_smart_attempt}/{_smart_max}: model stopped mid-task, nudging", file=sys.stderr)
+                    print(f"[{self._session_id}] [smart-continue] attempt {_smart_attempt}/{_smart_max}: model stopped mid-task (prior_ctx={_has_prior_tool_ctx} text_tools={_looks_like_tools}), nudging", file=sys.stderr)
                     try:
                         retry_upstream = urllib.request.urlopen(nudge_req, timeout=_upstream_timeout(body, True))
                         collected_events = []
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)

translate-proxy.py

@@ -1469,6 +1469,53 @@ _CROF_ADAPTIVE = {
     "min_keep_recent": 6,
 }
 
+_model_max_tokens = {}
+_model_max_tokens_lock = threading.Lock()
+
+def _estimate_tokens(item):
+    if not isinstance(item, dict):
+        return 4
+    t = item.get("type", "")
+    if t == "message":
+        content = item.get("content", "")
+        if isinstance(content, str):
+            return max(4, len(content) // 4)
+        elif isinstance(content, list):
+            total = 4
+            for part in content:
+                pt = part.get("type", "")
+                if pt in ("input_text", "output_text"):
+                    total += max(4, len(part.get("text", "")) // 4)
+                elif pt == "input_image":
+                    total += 800
+                elif pt in ("function_call",):
+                    total += max(20, len(part.get("arguments", "{}")) // 2)
+                elif pt == "function_call_output":
+                    total += max(8, len(part.get("output", "")) // 4)
+            return total
+    elif t in ("function_call_output",):
+        return max(8, len(item.get("output", "")) // 4)
+    elif t == "function_call":
+        return max(20, len(item.get("arguments", "{}")) // 2)
+    return 4
+
+def _estimate_input_tokens(input_data):
+    if not isinstance(input_data, list):
+        return 0
+    return sum(_estimate_tokens(i) for i in input_data)
+
+def _get_model_max_tokens(model):
+    with _model_max_tokens_lock:
+        return _model_max_tokens.get(model)
+
+def _set_model_max_tokens(model, tokens):
+    if model and tokens:
+        with _model_max_tokens_lock:
+            existing = _model_max_tokens.get(model)
+            if existing is None or tokens < existing:
+                _model_max_tokens[model] = tokens
+                print(f"[ctx-limit] learned {model} max ~{tokens} tokens", file=sys.stderr)
+
 _BGP_STATS_PATH = os.path.join(_LOG_DIR, "bgp-route-stats.json")
 _bgp_stats_lock = threading.Lock()
 
@@ -1534,8 +1581,6 @@ def _sorted_bgp_routes():
     return sorted(BGP_ROUTES, key=lambda r: _score_route(r, stats))
 
 def _crof_record(model, n_items, success):
-    if TARGET_URL and "crof.ai" not in TARGET_URL:
-        return
     if not isinstance(n_items, int) or n_items < 1:
         return
     entry = {"model": model, "items": n_items, "ok": success}
@@ -1561,20 +1606,36 @@ def _crof_record(model, n_items, success):
             global_limit = v["limit"]
     _CROF_ADAPTIVE["global_item_limit"] = global_limit
 
-    if TARGET_URL and "crof.ai" in TARGET_URL:
+    print(f"[crof-adaptive] model={model} items={n_items} {'OK' if success else 'FAIL'} -> limit={ml.get('limit',30)} global={global_limit}", file=sys.stderr)
-        print(f"[crof-adaptive] model={model} items={n_items} {'OK' if success else 'FAIL'} -> limit={ml.get('limit',30)} global={global_limit}", file=sys.stderr)
 
 def _crof_item_limit(model):
     ml = _CROF_ADAPTIVE["model_limits"].get(model, {})
     per_model = ml.get("limit", 30)
     return min(per_model, _CROF_ADAPTIVE["global_item_limit"])
 
-def _crof_compact_for_retry(input_data, model):
+def _crof_compact_for_retry(input_data, model, aggression=0):
     limit = _crof_item_limit(model)
-    if not isinstance(input_data, list) or len(input_data) <= limit:
+    if not isinstance(input_data, list) or len(input_data) < 2:
+        return input_data
+
+    max_tok = _get_model_max_tokens(model)
+    est = _estimate_input_tokens(input_data)
+    over_item_limit = len(input_data) > limit
+    over_token_limit = max_tok and est >= max_tok * 0.9
+
+    if not over_item_limit and not over_token_limit:
         return input_data
 
     keep = max(_CROF_ADAPTIVE["min_keep_recent"], limit // 3)
+    if over_token_limit:
+        ratio = est / max_tok
+        if aggression >= 1 or ratio > 1.5:
+            keep = max(2, _CROF_ADAPTIVE["min_keep_recent"] // 2)
+        elif ratio > 1.2:
+            keep = max(3, keep // 2)
+        print(f"[ctx-limit] model={model} est={est}tok max={max_tok}tok ratio={ratio:.2f} -> keep={keep}", file=sys.stderr)
+    elif over_item_limit:
+        keep = max(keep, 6)
     head_end = 0
     for i, item in enumerate(input_data):
         t = item.get("type")
@@ -1607,8 +1668,7 @@ def _crof_compact_for_retry(input_data, model):
         summary_lines.append(_item_summary(item, max_len=120))
 
     summary_msg = {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "\n".join(summary_lines)}]}
-    if TARGET_URL and "crof.ai" in TARGET_URL:
+    print(f"[crof-adaptive] RETRY compact: {len(input_data)} -> {len(head)+1+len(tail)} (limit={limit}, keep={len(tail)}, agg={aggression})", file=sys.stderr)
-        print(f"[crof-adaptive] RETRY compact: {len(input_data)} -> {len(head)+1+len(tail)} (limit={limit}, keep={len(tail)})", file=sys.stderr)
     return head + [summary_msg] + tail
 
 def _item_summary(item, max_len=200):
@@ -2051,6 +2111,18 @@ def synthesize_tool_results_for_chat(input_items):
 def has_function_call_output(input_items):
     return isinstance(input_items, list) and any(i.get("type") == "function_call_output" for i in input_items)
 
+_TOOL_CALL_TEXT_PATTERNS = re.compile(
+    r'(?:^|\n)[\s•\-\*]*\(?'
+    r'(?:exec_command|write_to_file|exec_bash|bash|run_command|shell|edit_file|read_file|search_files|list_files)'
+    r'[\s:]',
+    re.I | re.MULTILINE
+)
+
+def _text_looks_like_tool_calls(text):
+    if not text or len(text) < 6:
+        return False
+    return bool(_TOOL_CALL_TEXT_PATTERNS.search(text))
+
 # ═══════════════════════════════════════════════════════════════════
 # Log redaction
 # ═══════════════════════════════════════════════════════════════════
@@ -2233,9 +2305,14 @@ def _normalize_tool_args(raw_args):
     except json.JSONDecodeError:
         return raw_args
 
-_XML_TC_RE = re.compile(r'<tool_call>(\w+)(.*?)</tool_call>', re.DOTALL)
+_XML_TC_RE = re.compile(r'exec_command(.*?)</invoke>', re.DOTALL)
 _XML_ARG_VALUE_RE = re.compile(r'</?arg_value>\s*')
 
+_PAREN_TC_RE = re.compile(
+    r'(?:^|[\n•\-\*]\s*)\(\s*(exec_command|write_to_file|exec_bash|bash|run_command|shell|edit_file|read_file|search_files|list_files)\b\s*(.*?)\)',
+    re.DOTALL | re.I
+)
+
 def _extract_xml_tool_calls(text):
     if not text:
         return []
@@ -2262,6 +2339,68 @@ def _extract_xml_tool_calls(text):
         results.append({"name": name, "args": args_str, "call_id": f"xml_{len(results)}"})
     return results
 
+_NON_VISION_MODEL_PATTERNS = re.compile(
+    r'\b(deepseek|glm|mixtral|llama\b(?!.*vision)|command|dbrx|qwen\b(?!.*vl)|phi-?3(?!.*vision))',
+    re.I
+)
+
+_vision_fail_cache = set()
+_vision_fail_lock = threading.Lock()
+
+def _model_supports_vision(model):
+    if not model:
+        return True
+    with _vision_fail_lock:
+        if model in _vision_fail_cache:
+            return False
+    if _NON_VISION_MODEL_PATTERNS.search(model):
+        return False
+    return True
+
+def _mark_vision_fail(model):
+    if model:
+        with _vision_fail_lock:
+            _vision_fail_cache.add(model)
+
+def _strip_images_from_input(input_data, model):
+    if not isinstance(input_data, list) or _model_supports_vision(model):
+        return input_data
+    modified = False
+    result = []
+    for item in input_data:
+        if item.get("type") != "message":
+            result.append(item)
+            continue
+        content = item.get("content", [])
+        if isinstance(content, str):
+            result.append(item)
+            continue
+        new_content = []
+        has_img = False
+        for part in content:
+            if isinstance(part, str):
+                new_content.append(part)
+                continue
+            pt = part.get("type", "")
+            if pt in ("input_image", "image_url"):
+                if not has_img:
+                    fname = part.get("image_url", {}).get("url", part.get("url", "image.png"))
+                    if fname.startswith("data:"):
+                        fname = "screenshot.png"
+                    new_content.append({"type": "output_text", "text": f"[User attached image: {fname} — this model does not support vision]"})
+                    has_img = True
+                    modified = True
+            else:
+                new_content.append(part)
+        if modified:
+            result.append({**item, "content": new_content})
+        else:
+            result.append(item)
+    if modified:
+        print(f"[vision-filter] stripped {sum(1 for i in input_data if i.get('type')=='message' and any(c.get('type') in ('input_image','image_url') for c in (i.get('content') or []) if isinstance(c,dict)))} images for model={model}", file=sys.stderr)
+        return result
+    return input_data
+
 def oa_input_to_messages(input_data):
     msgs = []
     tool_name_by_id = {}
@@ -4889,12 +5028,25 @@ class Handler(http.server.BaseHTTPRequestHandler):
             body["input"] = input_data
 
         crof_limit = _crof_item_limit(model)
-        _crof_eligible = TARGET_URL and "crof.ai" in TARGET_URL
+        _crof_eligible = True
-        if _crof_eligible and not compacted and isinstance(input_data, list) and len(input_data) > crof_limit:
+        if _crof_eligible and not compacted and isinstance(input_data, list):
-            print(f"[crof-adaptive] proactive compact: {len(input_data)} items > limit {crof_limit}", file=sys.stderr)
+            _needs_compact = len(input_data) > crof_limit
-            input_data = _crof_compact_for_retry(input_data, model)
+            max_tok = _get_model_max_tokens(model)
-            body = dict(body)
+            est_tok = _estimate_input_tokens(input_data) if max_tok else 0
-            body["input"] = input_data
+            if not _needs_compact and max_tok and est_tok > max_tok * 0.8:
+                _needs_compact = True
+            if _needs_compact:
+                _agg = 0
+                if max_tok and est_tok > max_tok:
+                    _agg = 1
+                print(f"[crof-adaptive] proactive compact: {len(input_data)} items, est={est_tok}tok max={max_tok}tok agg={_agg}", file=sys.stderr)
+                input_data = _crof_compact_for_retry(input_data, model, aggression=_agg)
+                body = dict(body)
+                body["input"] = input_data
+
+        # Strip images for non-vision models
+        input_data = _strip_images_from_input(input_data, model)
+        body["input"] = input_data
 
         messages = oa_input_to_messages(input_data)
         messages = _inject_stored_reasoning(messages)
@@ -4927,14 +5079,19 @@ class Handler(http.server.BaseHTTPRequestHandler):
                 except urllib.error.HTTPError as e:
                     err_body = e.read().decode()
                     if "context_length_exceeded" in err_body and attempt < max_retries:
-                        print(f"[{self._session_id}] context_length_exceeded (attempt {attempt+1}/{max_retries}), retrying with extreme compaction!", file=sys.stderr)
+                        import re as _re
+                        _tok_m = _re.search(r'~?(\d+)\s*tokens', err_body)
+                        if _tok_m:
+                            _set_model_max_tokens(model, int(_tok_m.group(1)))
+                        print(f"[{self._session_id}] context_length_exceeded (attempt {attempt+1}/{max_retries}), retrying with compaction (agg={attempt})!", file=sys.stderr)
                         policy = provider_policy()
                         if isinstance(input_data, list):
-                            print(f"[{self._session_id}] applying extreme compaction to {len(input_data)} items", file=sys.stderr)
+                            est = _estimate_input_tokens(input_data)
-                            input_data = _crof_compact_for_retry(input_data, model)
+                            print(f"[{self._session_id}] applying compaction to {len(input_data)} items ~{est}tok", file=sys.stderr)
+                            input_data = _crof_compact_for_retry(input_data, model, aggression=attempt)
                             body = dict(body)
                             body["input"] = input_data
-                            messages = oa_input_to_messages(input_data)
+                            messages = oa_input_to_messages(_strip_images_from_input(input_data, model))
                             messages = _inject_stored_reasoning(messages)
                             instructions = body.get("instructions", "").strip()
                             if instructions:
@@ -5725,9 +5882,11 @@ class Handler(http.server.BaseHTTPRequestHandler):
             last_status = None
             finish_reason = None
             has_content = False
+            has_message = False
+            has_tool_call = False
 
             def _observe_event(event):
-                nonlocal last_resp_id, last_output, last_status, finish_reason, has_content
+                nonlocal last_resp_id, last_output, last_status, finish_reason, has_content, has_message, has_tool_call
                 for line in event.strip().split("\n"):
                     if line.startswith("data: "):
                         try:
@@ -5737,7 +5896,9 @@ class Handler(http.server.BaseHTTPRequestHandler):
                                 last_output = d.get("response", {}).get("output", [])
                                 last_status = d.get("response", {}).get("status")
                                 finish_reason = "length" if last_status == "incomplete" else "stop"
-                                has_content = any(o.get("type") == "message" for o in (last_output or []))
+                                has_tool_call = any(o.get("type") == "function_call" for o in (last_output or []))
+                                has_message = any(o.get("type") == "message" for o in (last_output or []))
+                                has_content = has_message or has_tool_call
                         except Exception:
                             pass
 
@@ -5749,7 +5910,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         break
                     collected_events.append(event)
                     _observe_event(event)
-                print(f"[{self._session_id}] stream ended: events={len(collected_events)} finish={finish_reason} has_content={has_content} elapsed={time.time()-t0:.1f}s", file=sys.stderr)
+                print(f"[{self._session_id}] stream ended: events={len(collected_events)} finish={finish_reason} has_content={has_content} has_message={has_message} has_tool_call={has_tool_call} elapsed={time.time()-t0:.1f}s", file=sys.stderr)
             except (ConnectionResetError, BrokenPipeError, ConnectionAbortedError):
                 print("[translate-proxy] client disconnected during stream", file=sys.stderr)
                 _crof_record(model, n_items, False)
@@ -5805,6 +5966,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)
@@ -5813,7 +5976,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         print(f"[provider-sensor] synthetic retry failed: {e}", file=sys.stderr)
 
             # Auto-retry on finish_reason=length with no content due to too much context.
-            if finish_reason == "length" and not has_content and isinstance(input_data, list) and len(input_data) > 5 and TARGET_URL and "crof.ai" in TARGET_URL:
+            if finish_reason == "length" and not has_content and isinstance(input_data, list) and len(input_data) > 5:
                 print(f"[crof-adaptive] RETRY: finish_reason=length with no content, compacting {n_items} items", file=sys.stderr)
                 new_input = _crof_compact_for_retry(input_data, model)
                 if len(new_input) < len(input_data):
@@ -5836,6 +5999,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)
@@ -5943,9 +6108,17 @@ class Handler(http.server.BaseHTTPRequestHandler):
                 _smart_attempt = 0
                 while _smart_attempt < _smart_max:
                     _has_tool_calls_in_output = any(o.get("type") == "function_call" for o in (last_output or []))
+                    last_text = ""
+                    for o in (last_output or []):
+                        if o.get("type") == "message":
+                            for c in (o.get("content") or []):
+                                if isinstance(c, dict) and c.get("type") == "output_text":
+                                    last_text += c.get("text", "")
+                    _looks_like_tools = _text_looks_like_tool_calls(last_text)
+                    _has_prior_tool_ctx = has_function_call_output(input_data)
                     if not (finish_reason == "stop" and has_content and not _has_tool_calls_in_output
                             and isinstance(input_data, list) and len(input_data) >= 3
-                            and has_function_call_output(input_data)):
+                            and (_has_prior_tool_ctx or _looks_like_tools)):
                         break
                     _smart_attempt += 1
                     _nudges = [
@@ -5954,12 +6127,6 @@ class Handler(http.server.BaseHTTPRequestHandler):
                     ]
                     nudge_text = _nudges[min(_smart_attempt - 1, len(_nudges) - 1)]
                     # Try extracting XML tool calls from text as fallback before nudging
-                    last_text = ""
-                    for o in (last_output or []):
-                        if o.get("type") == "message":
-                            for c in (o.get("content") or []):
-                                if isinstance(c, dict) and c.get("type") == "output_text":
-                                    last_text += c.get("text", "")
                     xml_fc = _extract_xml_tool_calls(last_text)
                     if xml_fc:
                         print(f"[{self._session_id}] [smart-continue] extracted {len(xml_fc)} XML tool calls from text, injecting and retrying", file=sys.stderr)
@@ -5979,6 +6146,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
                             last_resp_id = last_output = last_status = None
                             finish_reason = None
                             has_content = False
+                            has_message = False
+                            has_tool_call = False
                             for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                                 collected_events.append(event)
                                 _observe_event(event)
@@ -5988,19 +6157,21 @@ class Handler(http.server.BaseHTTPRequestHandler):
                             print(f"[{self._session_id}] [smart-continue] XML injection retry failed: {e}", file=sys.stderr)
                             break
                     _nudge_msg = {"role": "user", "content": nudge_text}
-                    nudge_messages = oa_input_to_messages(input_data) + [_nudge_msg]
+                    nudge_messages = oa_input_to_messages(_strip_images_from_input(input_data, model)) + [_nudge_msg]
                     instructions = body.get("instructions", "").strip()
                     if instructions:
                         nudge_messages.insert(0, {"role": "system", "content": instructions})
                     nudge_chat_body = self._build_chat_body(model, nudge_messages, body, stream)
                     nudge_req = urllib.request.Request(target, data=json.dumps(nudge_chat_body).encode(), headers=fwd)
-                    print(f"[{self._session_id}] [smart-continue] attempt {_smart_attempt}/{_smart_max}: model stopped mid-task, nudging", file=sys.stderr)
+                    print(f"[{self._session_id}] [smart-continue] attempt {_smart_attempt}/{_smart_max}: model stopped mid-task (prior_ctx={_has_prior_tool_ctx} text_tools={_looks_like_tools}), nudging", file=sys.stderr)
                     try:
                         retry_upstream = urllib.request.urlopen(nudge_req, timeout=_upstream_timeout(body, True))
                         collected_events = []
                         last_resp_id = last_output = last_status = None
                         finish_reason = None
                         has_content = False
+                        has_message = False
+                        has_tool_call = False
                         for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
                             collected_events.append(event)
                             _observe_event(event)